Causal Networks: Semantics and Expressiveness*
|
|
- Katrina Miles
- 6 years ago
- Views:
Transcription
1 Causal Networks: Semantics and Expressiveness* Thomas Verina & Judea Pearl Cognitive Systems Laboratory, Computer Science Department University of California Los Angeles, CA ABSTRACT Dependency knowledge of the fonn " x is independent of y once z is known" invariably obeys a set of four axioms defining semi-graphoids, examples of which are probabilistic and database dependencies. Such knowledge can often be stored efficiently in graphical structures, using either undirected graphs and directed acyclic graphs (DAGs). This paper shows that, if we construct a DAG from any causal input list of a semi-graphoid, then a graphical criterion called d-separation is a sound rule for reading independencies from the DAG, i.e., it produces only valid assertions of conditional independence. The rule is extended to include DAGs with functio al dependencies. NTRODUCTON Dependency knowledge is useful in several areas of research, for example in database design it is useful to reason about embedded-multivalued-dependence (EMVD) of attributes [Fagin, 1977] and in decision analysis and expert systems design it is useful to reason about probabilistic independence of variables [Howard and Matheson, 1981], [Pearl, 1988). These represent two formalizations of the intuitive relation "knowing Z renders X and Y independent" which shall be denoted as (X, Z, Y). This relation would naturally have different properties in different applications, but it is interesting to note that most sensible definitions of this relation share four common properties listed below: symmetry decomposition weak union contraction (X, Z, Y) <=> (Y, Z, X) (X, Z, YW) => (X, Z, Y) /(X, Z, YW) => (X, 2Y, W) (X, ZY, W) & (X, Z, Y) => (X, Z, YW) (l.a) (l.b) (l.c) (. d) where X, Y and Z represent three disjoint subsets of objects (e.g. variables, attributes). t is known that every EMVD relation obeys the four properties listed above, as well as many other properties. (The notation (X, Z, Y) is equivalent to the standard EMVD notation Z X Y). Probabilistic dependencies also obey these four properties [Dawid, 1979), and it has been conjectured that they are, in fact, complete [Pearl and Paz, 1985), namely, that any other property of probabilistic independence is a logical consequence of the four. Three place relations which obey these four properties are called semi-graphoids. * This work was partially supported by the National Science Foundation Grant #R "Graphoids: A Computer Representation for Dependencies and Relevance in Automated Reasoning (Computer nformation Science)." 352
2 1'. For probability distributions containing strictly positive probabilities, the independence relation has a fifth independent property: intersection (X, Zf, W) & (X, ZW, Y) :::) (X, Z, YW) (2) This property along with the four of semi-graphoids define the class of graphoids (also conjectured to be complete for non-extreme probabilities [Pearl and Paz, 1985]). A naive approach for representing a dependency model, i.e., particular instance of a dependency relation, would be to enumerate all triplets (X, Z, Y) for which /(X, Z, Y) holds. This could result in exponential space since the relation ranges over subsets of objects. The use of graphs as a representation of dependency models is appealing in three ways; first the graph has an intuitive conceptual meaning, second, it is an efficient representation in tenns of time and space [Pearl and Venna, 1987], third there are efficient methods which utilize the infonnation when it is organized this way [Pearl, 1982, 1985, 1986, 1988], [Shachter, 1985, 1988]. UNDRECTED GRAPHS The meaning of a particular undirected graph is straight forward, each node in the graph represents a variable, and a link in the graph means that the two variables are directly dependent. Under this meaning, a set of nodes Z would separate two other sets X and Y, if and only if every path between a node in X and a node in Y passes through Z. This representation can fully represent only a small set of dependency models defined by the following properties [Pearl and Paz, 1985]: symmetry decomposition strong union intersection transitivity (X, Z, Y) <=> (Y, Z, X) (X, Z, YW) :::) (X, Z, Y) (X, Z, Y) :::) (X, ZW, Y) (X, Zf, W) & (X, ZW, Y) ::) (X, Z, YW) (X, Z, Y) :::) (X, Z, y) or (Y, Z, y) 'r/y X u Y u Z t is not always necessary to have an exact graphical representation of a dependency model, in fact an efficient approximation called an 1-map is often preferred to an inefficient perfect map. A representation R of a dependency model M is an 1-map if every independence represented in R implies a valid independence in M. Thus, R may not contain au independencies of M, but the ones it does ontain are correct There is an algorithm which finds the most representative 1-map for any graphoid [Pearl and Paz, 1985]. Since probabilistic independence over positive probabilities constitutes a graphoid, there is always a unique edge-minimal undirected graph which is an 1-map of any non-extreme probabilistic distribution P (i.e., no edge can be removed without destroying the -mapness of the graph). This is not the case for EMVD relations nor for probabilistic distributions that exclude certain combinations of events; in these cases there is no unique edge-minimal -map for a given model, and, moreover, there is no effective method of constructing even one minimal -map. (3.a) (3.b) (3.c) (3.d) (3.e) 353
3 DRECTED-ACYCLC GRAPHS (DAGS) The dependency model represented by a particular DAG has a simple causal interpretation; each node represents a variable and there is a directed arc from one node to another if the first is a direct cause of the second. Under this interpretation, graph-separation is not as straight forward as before since two unrelated causes of a symptom may become related once the symptom is observed [Pearl, 1986]. Thus a set of nodes Z is defined to d-separate two other sets X and Y if and only if every adjacency path from a node in X to a node in Y is rendered inactive by Z. An adjacency path is a path which follows arcs ignoring their directionality; one is rendered inactive by a set of nodes Z if and only if either there is a headto-head node along the path which is. not in Z and none of its descendents are in Z or some node along the path is not head-to-head but is in Z. A node along the path is head-to-head if the node before it and after it along the path both point to it in the graph. One node is a descendent of another if there is a directed path from the latter to the former. There is a procedure which produces a minimal 1-map of any dependency model which is a semi-graphoid. t employs an algorithm which takes a stratified protocol (also called causal input list [Geiger and Pearl, 1988]) of a dependency model and produces a perfect map of its semi-graphoid closure. A stratified protocol of a dependency model contains two things: an ordering of the variables, and a function that assigns a tail boundary to each variable x. A tail boundary of a variable x is any set of lesser variables (with respect to the ordering) rendering x independent of all other lesser variables. A unique DAG can be generated from each stratified protocol by associating the set of direct parents of any node x in the DAG with the tail boundary of the variable x in the protocol. An equivalent specification of a stratified protocol is an ordered list of triplets of the form (n, B, R), one triplet for each variable in the model, where the set B is the tail boundary of the variable n and R is a set containing all other lesser variables. For a particular dependency model over n variables there are n! orderings, and for each ordering ll there can be up to T 2k- = 2" ( n- )2 different sets of tail boundaries since, in the worst case, every subk= set of lesser variables could be a boundary. Thus, there can be as many as n! 2" ( n- l)/z stratified protocols. But if the dependency model posses a perfect map in DAGs, then one of the protocols is guaranteed to generate it by the following theorem. Theorem 1: f M is a dependency model which can be perfectly represented by some DAG D, then there is a stratified protocol La which generates D. Proof: Let D be a DAG which perfectly represents M. Since D is a directed acyclic graph it imposes a partial order cp on the variables of M. Let a be any total ordering consistent with cp (i.e. a< ' b =>a <a b). For any node n in D, the set of its parents P (n) constitutes a tail boundary with respect to the ordering a, thus the pair La= (a, P (n)) is a stratified protocol of M, and this is the very protocol which will generate D. QED. Although it is i>ossible to find a perfect map when it exists, testing for existence may be intractable. t is practical, however, to find a minimal 1-map, and the next theorem shows that stratified protocols can be used to generate 1-maps of any semi-graphoid, not necessarily those possessing perfect maps in DAGs. 354 ',,
4 ,,,. J.. Theorem 2: f M is a semi-graphoid, and La is any stratified protocol of M, then the DAG generated by Le is an-map of M. Proof: nduct on the number of variables in the semi-graphoid. For semi-graphoids of one variable it is obvious that the DAG generated is an -map. Suppose for semi-graphoids with fewer than k variables that the DAG is also an -map. Let M have k variables, n be the last variable in the ordering a, M- n be the semi-graphoid formed by removing n and all triplets involving n from M and G-n be the DAG formed by removing n and all its incident links from G. Since n is the last variable in the ordering, it cannot appear in any of boundaries of L 9, and thus L 9-n can be defined to contain only the first n -1 variables and boundaries of Le and still be a stratified protocol of M -n. n fact the DAG generated from La -n is G-n. Since M-n has k-1 variables, G-n is an -map of it Let Ma be the dependency model corresponding to the DAG G, and Ma-n correspond to G-n, (i.e. Ma contains all d-separated triplets of G). G is an-map of M if and only if Ma M. Each triplet T of Ma falls into one of four categories; either the variable n does not appear in Tor it appears in the first, second or third entry oft. These will be treated separately as cases 1, 2, 3 and 4, respectively. ease-l: f n does not appear in T then T must equal (X, Z, Y) with X, Y and Z three disjoint subsets of variables, none of which contain n. Since Tis in Ma it must also be in M a -n for if it were not then there would be an active path in G- n between a node in X and a node in Y when Z is instantiated. But if this path is active in G -n then it must also be active in G since the addition of nodes and links can not deactivate a path. Since G -n is an -map of M- n, T must also be an element of it, but M -n is a subset of M,so Tis inm ease-l: f n appears in the first entry of the triplet, then T = (Xn, Z, Y) with the same constraints on X, Y and Z as in ease-l. Let (n, B, R) be the last triple in L9, Bx. By, Bz and B 0 be a partitioning of B and Rx. Ry, Rz and R o be a partitioning of R such that X = Bx u Rx, Y =By u Ry and Z = Bz u Rz as in figure 1. Figure 1 By the method of construction, there is an arrow from every node in B to n, but since (Xn, z. Y) is in Ma every path from a node in Y to n must be deactivated by Z so By must be empty or else there would be a direct link from Y to n (see figure 2a). The last triplet in Le can now be written as (n, BxBoBz,RxRzYRo). Since X =Bx urx. Y=Ry and M is a semi-graphoid it follows (from (1.b) and (l.c)) that (n, XB 0 Z, Y) e M. 355 Ro
5 Figure 2a Figure 2b. Since there is an arrow from every node in B o to n and n is separated from Y given Z in G, B 0 must also be d-separated from Y given Z in G for if it were connected there would be a path from a node in Y to a node in B 0 which was active given Z. But there is an arrow from every node in B 0 to n, thus, such a path would also connect the node in Y to n, and Y would no longer be separated from n given Z (see figure 2b). Since Y is separated from both B 0 and X given Z in the DAG G it is separated from their union, so (XB0,Z, Y)e M0. Since n is not in this triplet, the argument of ease-l above implies that (XB 0, Z, Y) e M. Since (n, XB oz. Y) e M and M is a semi-graphoid it follows (using (l.b) and (l.d)) that T=(Xn, Z, Y)e M case-3: f n appears in the second entry then T = (X, Zn, Y) with the same constraints on X, Y and Z as in ease-l. Also let B, R, etc. be defined as in case-2, thus (n, BxBrBzBo, RxRrRzRo)e M since it is the last triplet in L9 n this case, either By is empty and B 0 is separated from Y given ZinG, or Bx is empty and B 0 is separated from X given Z in G since if neither were the case, then there would be a path from a node in Y which would be active given Z and would end pointing at n, and there would be a similar path from a node in X to n. But this means that there would be a path from a node in X to a node in Y which would be active given Z and n since the path is head-to-head at n (see figure 3a and 3b). But there can be no such path since by assumption (X, Zn, Y) e M0. Without loss of generality, assume that By = 0 and (Bo,Z, Y)e Mo.. Figure 3a Figure 3b X and Y must be separated in G given only Z for if they were not then there would be a path between them 356
6 ,, which would be active given Z. Since they are separated given Z and n, this path would have to be deactivated by n, but since there are only arrows pointing at n it can only activate paths by being instantiated (see figure 3b). Thus, there can be no such path and X andy must be separated given Z ing. Since Yis separated from both X and B 0 given Z in G it follows that (XB o. Z, Y) e Ma and by the argument of ease-l above (XB 0, Z, Y) em. Since By is empty it follows, as in case-2, that (n, XB 0Z, Y) em. Further, M is a semi-graphoid so, applying (l.d) to (XB 0, Z, Y) e M and (n, XB 0Z, Y) e M yields (nxb0, z, Y)e M and using (l.b) and (l.c) it follows that T=(X, Zn, Y) em. case-4: f n appears in the third entry, then by symmetry the triplet T is equivalent to one with n in the first entry, and the argument of case-2 above shows that T em. QED.. Corollary 1: f Leis any stratified protocol ofsome dependency model M, the DAG generated from La is a perfect map of the semi-graphoid closure of L9 n other words, a triplet is d-separated in the DAG if and only if it can be derived from the triplets of La using the four axioms in (1). Proof: By the previous theorem, the DAG is an 1-map of the closure, and it remains to show that the closure is an 1-map of the DAG. Since every DAG dependency model is a semi-graphoid, the DAG closure of La contains the semi-graphoid closure of it, thus, it suffices to show that the DAG dependency model MG contains L8. f (n, B, R) is a triplet in La then n is separated from R given B in the DAG. for if not then there would be a path from a node in R to n which is active given B. But since every link into n is from B the path must lead out of n into some node which was placed after n. Since every node in R was placed before n, this path cannot be directed and must contain a head-to-head node at some node which was placed after n. But this path is deactivated by B since it contains no nodes placed after n, and thus, B would separate n from R in the graph. QED. Corollary 2: f each tail boundary in La is minimal, the resulting DAG is a minimal 1-map of M. Theorem 2 and its corollaries together imply that d-separation is sound and complete for the extraction of independence infonnation from DAGs with respect to their stratified protocols. That is, a conclusion can be read from the graph using d-separation if and only if it follows from application of the semi-graphoid axioms to the stratified protocol. n bayesian networks [Pearl, 1986), for example, any independence which can be read from the graph via d-separation is sound with respect to the probability distribution that it represents. Since the axioms of semi-graphoids have not been shown to be complete for the class of probabilistic dependencies, corollary 1 is not enough to ensure that d-separation will identify more independencies than any other sound criterion. The latter has been shown in [Geiger and Pearl, 1988]. The last theorem, which is of a theoretical nature, states that it is possible to force any particular independence of a semi-graphoid to be represented in an 1-map. Theorem 3: f M is any semi-graphoid then the set of DAGs generated from all stratified protocols of M is a perfect map of M if the criterion for separation is that d-separation must exist in one of the DAGs. Proof: f there is a separation in one of the DAGs then the corresponding independence must hold in M since theorem 2 states that each of the DAGs is an 1-map of M, thus the set is also an 1-map. t remains to show that M is an 1-map of the set of DAGs. LetT= (X, Z, Y) be any triplet in M and X= {x1, 0 0 0,Xn} '
7 The triplets r * = { (x;, x 1 x; _1 Z, Y) 1 S i S n } must also be in M since they are implied by T using the weak union axiom of semi-graphoids. Furthennore Tis in the semi-graphoid closure of r * since the triplets imply T by use of the contraction axiom. Thus any protocol containing the triplets r * would generate a DAG containing T. Such a protocol need only have an ordering a such that the variables of Y and Z are less than those of X which are less than any other variables and that the variables of X are ordered such that x; <e x i if and only if i < j. The DAG generated by this protocol is in the set of DAGs and therefore the separation holds in the set. QED. Since there is an effective algorithm for generating an-map DAG for any semi-graphoid, DAGs would be a useful means of representing EMVD relations as well as probabilistic independence relations. Furthermore if the particular dependency model is stated as a stratified protocol then it can be perfectly represented by a DAG. FUNCTONAL DEPENDENCES The ability to represent functional dependencies would be a powerful extension from the point of view of the designer. These dependencies may easily be represented by the introduction of deterministic nodes which would correspond to the deterministic variables [Shachter, 1988]. Graphs which contain deterministic nodes represent more information than d-separation is able to extract; but a simple extension of d-separation, called D-separation, is both sound and complete with respect to the input protocol under both probabilistic inference and semi-graphoid inference [Geiger and Venna, 1988]. D-separation is very similar to d-separation, only differing in that a path is rendered inactive by a set of nodes Z under D-separation just in the case that it would be inactive under d-separation plus the case there is a node on the path which is determined by Z. CONCLUSONS This paper shows that the criteria of d-separation and D-separation are sound. They provide a reliable and efficient method for extracting independence infonnation from DAGs. This information may be used explicitly, for example to help guide a complex reasoning system, or implicitly as in bayesian propagation [Pearl, 1986] and the evaluations of nfluence Diagrams [Shachter, 1985, 1988]. The criteria also provide a sound theoretical basis for the analysis of the properties of these graphical representations. For example, the validity of graphical manipulations such as arc reversal and node removal [Howard and Matheson, 1981] [Shachter 1985, 1988][Smith, 1987] can now be affirmed on solid theoretical foundations. ACKNOWLEDGMENT We thank Dan Geiger and Azaria Paz for many valuable discussions, and James Smith for checking the proof of Theorem
8 REFERENCES Dawid A.P., Conditional ndependence in Statistical Theory, J.R. Statist.B., vol. 41, No. 1, pp; Fagin R., Multivalued Dependencies and a New Form for Relational Databases, ACM Transactions on Database Systems, vol 2, No. 3, pp Geiger D. and Pearl J., On the Logic of Causal Models, UCLA Cognitive Systems Laboratory, Technical Report (R-112). n [Geiger and Pearl, 1988], this volume. Geiger D. and Verma T.S., Recognizing ndependence in nfluence Diagrams with Deterministic Nodes, UCLA Cognitive Systems Laboratory, Technical Report R-116, in preparation. Howard R.A. and Matheson J.E., nfluence Diagrams, in Principles and Applications of Decision Analysis, Menlo Park, Ca:Strategic Decision Group. Pearl J., Reverend Bayes on nference Engines: A Distributed Hierarchical Approach, Proc. of the Nat/. Conference on A, Pittsburgh, pp Pearl J., A Constraint Propagation Approach to Probabilistic Reasoning. Proc. of the First Workshop on Uncertainty ina/, Los Angeles, pp Pearl J., Fusion, Propagation and Structuring in Belief Networks, Artificial ntelligence, vol. 29, No 3, pp Pearl J., Probabilistic Reasoning in ntelligent Systems, San Mateo:Morgan-Kaufmann, Pearl J. and Paz A., GRAPHODS: A Graph-based Logic for Reasoning about Relevance Relations, UCLA Computer Science Department, Technical Report (R-53). Also in Advances in Artificial ntelligence-//, Edited by Du Boulay et al, Amsterdam:North-Holland Publishing Co Pearl J. and Verma T.S., The Logic of Representing Dependencies by Directed Graphs, Proc. of the 6th Natl. Conference on A (AAA/-87), Seattle, vol. 1, pp Shachter R., ntelligent Probabilistic nference, Proc. of the First Workshop on Uncertainty in A, Los Angeles, pp Shachter R., Probabilistic nference and nfluence Diagrams, to appear in Operations Research. Smith, J.Q., nfluence Diagrams for Bayesian Decision Analysis, Technical Report #100, Department of Statistics, University of Warwick, England. 359
CAUSAL NETWORKS: SEMANTICS AND EXPRESSIVENESS*
Uncertainty in Artificial Intelligence 4 R.D. Shachter, T.S. Levitt, L.N. Kanal, J.F. Lemmer (Editors) Elsevier Science Publishers B.V. (North-Holland), 1990 69 CAUSAL NETWORKS: SEMANTICS AND EXPRESSIVENESS*
More informationCONDITIONAL INDEPENDENCE AND ITS REPRESENTATIONS*
CONDITIONAL INDEPENDENCE AND ITS REPRESENTATIONS* JUDEA PEARL, DAN GEIGER, THOMAS VERMA This paper summarizes recent investigations into the nature of informational dependencies and their representations.
More informationA note on the pairwise Markov condition in directed Markov fields
TECHNICAL REPORT R-392 April 2012 A note on the pairwise Markov condition in directed Markov fields Judea Pearl University of California, Los Angeles Computer Science Department Los Angeles, CA, 90095-1596,
More informationCheng Soon Ong & Christian Walder. Canberra February June 2018
Cheng Soon Ong & Christian Walder Research Group and College of Engineering and Computer Science Canberra February June 2018 Outlines Overview Introduction Linear Algebra Probability Linear Regression
More informationMarkov Equivalence in Bayesian Networks
Markov Equivalence in Bayesian Networks Ildikó Flesch and eter Lucas Institute for Computing and Information Sciences University of Nijmegen Email: {ildiko,peterl}@cs.kun.nl Abstract robabilistic graphical
More informationA Transformational Characterization of Markov Equivalence for Directed Maximal Ancestral Graphs
A Transformational Characterization of Markov Equivalence for Directed Maximal Ancestral Graphs Jiji Zhang Philosophy Department Carnegie Mellon University Pittsburgh, PA 15213 jiji@andrew.cmu.edu Abstract
More informationThese notes present some properties of chordal graphs, a set of undirected graphs that are important for undirected graphical models.
Undirected Graphical Models: Chordal Graphs, Decomposable Graphs, Junction Trees, and Factorizations Peter Bartlett. October 2003. These notes present some properties of chordal graphs, a set of undirected
More informationNode Aggregation for Distributed Inference in Bayesian Networks
Node Aggregation for Distributed Inference in Bayesian Networks Kuo-Chu Chang and Robert Fung Advanced Decision Systmes 1500 Plymouth Street Mountain View, California 94043-1230 Abstract This study describes
More informationAutomated construction of sparse Bayesian networks from unstructured probabilistic models and domain information
Automated construction of sparse Bayesian networks from unstructured probabilistic models and domain information Sampath Srinivas Rockwell Science Center, Palo Alto Laboratory 444 High Street Palo Alto,
More informationGraphical Analysis of Value of Information in Decision Models
From: FLAIRS-01 Proceedings. Copyright 2001, AAAI (www.aaai.org). All rights reserved. Graphical Analysis of Value of Information in Decision Models Songsong Xu Kim-Leng Poh Department of lndustrial &
More informationFMA901F: Machine Learning Lecture 6: Graphical Models. Cristian Sminchisescu
FMA901F: Machine Learning Lecture 6: Graphical Models Cristian Sminchisescu Graphical Models Provide a simple way to visualize the structure of a probabilistic model and can be used to design and motivate
More informationResearch Article Structural Learning about Directed Acyclic Graphs from Multiple Databases
Abstract and Applied Analysis Volume 2012, Article ID 579543, 9 pages doi:10.1155/2012/579543 Research Article Structural Learning about Directed Acyclic Graphs from Multiple Databases Qiang Zhao School
More informationI I I I I I I I I I I I I I I I I I I
UPDATNG PROBABLTES N MULTPLY-CONNECTED BELEF NETWORKS H.J. Suennondt and G.F. Cooper Medical Computer Science Group Stanford University Stanford, California 94305-5479 This paper focuses on probability
More informationChapter 2 PRELIMINARIES. 1. Random variables and conditional independence
Chapter 2 PRELIMINARIES In this chapter the notation is presented and the basic concepts related to the Bayesian network formalism are treated. Towards the end of the chapter, we introduce the Bayesian
More informationThe strong chromatic number of a graph
The strong chromatic number of a graph Noga Alon Abstract It is shown that there is an absolute constant c with the following property: For any two graphs G 1 = (V, E 1 ) and G 2 = (V, E 2 ) on the same
More informationAv. Prof. Mello Moraes, 2231, , São Paulo, SP - Brazil
" Generalizing Variable Elimination in Bayesian Networks FABIO GAGLIARDI COZMAN Escola Politécnica, University of São Paulo Av Prof Mello Moraes, 31, 05508-900, São Paulo, SP - Brazil fgcozman@uspbr Abstract
More informationLecture 2 - Graph Theory Fundamentals - Reachability and Exploration 1
CME 305: Discrete Mathematics and Algorithms Instructor: Professor Aaron Sidford (sidford@stanford.edu) January 11, 2018 Lecture 2 - Graph Theory Fundamentals - Reachability and Exploration 1 In this lecture
More informationOn the Implication Problem for Probabilistic Conditional Independency
IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS PART A: SYSTEMS AND HUMANS, VOL. 30, NO. 6, NOVEMBER 2000 785 On the Implication Problem for Probabilistic Conditional Independency S. K. M. Wong, C.
More informationTesting Independencies in Bayesian Networks with i-separation
Proceedings of the Twenty-Ninth International Florida Artificial Intelligence Research Society Conference Testing Independencies in Bayesian Networks with i-separation Cory J. Butz butz@cs.uregina.ca University
More informationA Well-Behaved Algorithm for Simulating Dependence Structures of Bayesian Networks
A Well-Behaved Algorithm for Simulating Dependence Structures of Bayesian Networks Yang Xiang and Tristan Miller Department of Computer Science University of Regina Regina, Saskatchewan, Canada S4S 0A2
More informationProbabilistic Graphical Models
Overview of Part One Probabilistic Graphical Models Part One: Graphs and Markov Properties Christopher M. Bishop Graphs and probabilities Directed graphs Markov properties Undirected graphs Examples Microsoft
More informationThe Structure of Bull-Free Perfect Graphs
The Structure of Bull-Free Perfect Graphs Maria Chudnovsky and Irena Penev Columbia University, New York, NY 10027 USA May 18, 2012 Abstract The bull is a graph consisting of a triangle and two vertex-disjoint
More informationA Logical Notion of Conditional Independence: Properties and Applications
From: AAAI Technical Report FS-94-02. Compilation copyright 1994, AAAI (www.aaai.org). All rights reserved. A Logical Notion of Conditional Independence: Properties and Applications Adnan Darwiche Rockwell
More informationFaster Constraint Satisfaction Algorithms for Temporal Reasoning' (*)
TECHNICAL REPORT R-151 July, 1990 Faster Constraint Satisfaction Algorithms for Temporal Reasoning' (*) Itay Meiri and Judea Pearl Cognitive Systems Laboratory Computer Science Department University of
More informationTrees. 3. (Minimally Connected) G is connected and deleting any of its edges gives rise to a disconnected graph.
Trees 1 Introduction Trees are very special kind of (undirected) graphs. Formally speaking, a tree is a connected graph that is acyclic. 1 This definition has some drawbacks: given a graph it is not trivial
More informationComputer Vision Group Prof. Daniel Cremers. 4. Probabilistic Graphical Models Directed Models
Prof. Daniel Cremers 4. Probabilistic Graphical Models Directed Models The Bayes Filter (Rep.) (Bayes) (Markov) (Tot. prob.) (Markov) (Markov) 2 Graphical Representation (Rep.) We can describe the overall
More informationTopic 3: Propositions as types
Topic 3: Propositions as types May 18, 2014 Propositions as types We have seen that the main mathematical objects in a type theory are types. But remember that in conventional foundations, as based on
More informationA step towards the Bermond-Thomassen conjecture about disjoint cycles in digraphs
A step towards the Bermond-Thomassen conjecture about disjoint cycles in digraphs Nicolas Lichiardopol Attila Pór Jean-Sébastien Sereni Abstract In 1981, Bermond and Thomassen conjectured that every digraph
More informationThe External Network Problem
The External Network Problem Jan van den Heuvel and Matthew Johnson CDAM Research Report LSE-CDAM-2004-15 December 2004 Abstract The connectivity of a communications network can often be enhanced if the
More informationGraph Connectivity G G G
Graph Connectivity 1 Introduction We have seen that trees are minimally connected graphs, i.e., deleting any edge of the tree gives us a disconnected graph. What makes trees so susceptible to edge deletions?
More informationSTABILITY AND PARADOX IN ALGORITHMIC LOGIC
STABILITY AND PARADOX IN ALGORITHMIC LOGIC WAYNE AITKEN, JEFFREY A. BARRETT Abstract. Algorithmic logic is the logic of basic statements concerning algorithms and the algorithmic rules of deduction between
More informationGraphical Models. David M. Blei Columbia University. September 17, 2014
Graphical Models David M. Blei Columbia University September 17, 2014 These lecture notes follow the ideas in Chapter 2 of An Introduction to Probabilistic Graphical Models by Michael Jordan. In addition,
More informationGraphical Models. Pradeep Ravikumar Department of Computer Science The University of Texas at Austin
Graphical Models Pradeep Ravikumar Department of Computer Science The University of Texas at Austin Useful References Graphical models, exponential families, and variational inference. M. J. Wainwright
More informationReducing Directed Max Flow to Undirected Max Flow and Bipartite Matching
Reducing Directed Max Flow to Undirected Max Flow and Bipartite Matching Henry Lin Division of Computer Science University of California, Berkeley Berkeley, CA 94720 Email: henrylin@eecs.berkeley.edu Abstract
More informationComputer Vision Group Prof. Daniel Cremers. 4. Probabilistic Graphical Models Directed Models
Prof. Daniel Cremers 4. Probabilistic Graphical Models Directed Models The Bayes Filter (Rep.) (Bayes) (Markov) (Tot. prob.) (Markov) (Markov) 2 Graphical Representation (Rep.) We can describe the overall
More informationA Reduction of Conway s Thrackle Conjecture
A Reduction of Conway s Thrackle Conjecture Wei Li, Karen Daniels, and Konstantin Rybnikov Department of Computer Science and Department of Mathematical Sciences University of Massachusetts, Lowell 01854
More informationBelief propagation in a bucket-tree. Handouts, 275B Fall Rina Dechter. November 1, 2000
Belief propagation in a bucket-tree Handouts, 275B Fall-2000 Rina Dechter November 1, 2000 1 From bucket-elimination to tree-propagation The bucket-elimination algorithm, elim-bel, for belief updating
More informationUtilizing Device Behavior in Structure-Based Diagnosis
Utilizing Device Behavior in Structure-Based Diagnosis Adnan Darwiche Cognitive Systems Laboratory Department of Computer Science University of California Los Angeles, CA 90024 darwiche @cs. ucla. edu
More informationEscola Politécnica, University of São Paulo Av. Prof. Mello Moraes, 2231, , São Paulo, SP - Brazil
Generalizing Variable Elimination in Bayesian Networks FABIO GAGLIARDI COZMAN Escola Politécnica, University of São Paulo Av. Prof. Mello Moraes, 2231, 05508-900, São Paulo, SP - Brazil fgcozman@usp.br
More informationBuilding Classifiers using Bayesian Networks
Building Classifiers using Bayesian Networks Nir Friedman and Moises Goldszmidt 1997 Presented by Brian Collins and Lukas Seitlinger Paper Summary The Naive Bayes classifier has reasonable performance
More informationIntroduction to information theory and coding - Lecture 1 on Graphical models
Introduction to information theory and coding - Lecture 1 on Graphical models Louis Wehenkel Department of Electrical Engineering and Computer Science University of Liège Montefiore - Liège - October,
More informationarxiv:submit/ [math.co] 9 May 2011
arxiv:submit/0243374 [math.co] 9 May 2011 Connectivity and tree structure in finite graphs J. Carmesin R. Diestel F. Hundertmark M. Stein 6 May, 2011 Abstract We prove that, for every integer k 0, every
More informationI I I I I I I I I I I I I I I I I I I
Stochastic Simulation of Bayesian Belief Networks Homer L. Chin and Gregory F. Cooper Medical Computer Science Group Knowledge Systems Laboratory Medical School Office Building, Room 215 Stanford University
More informationLecture 4: Undirected Graphical Models
Lecture 4: Undirected Graphical Models Department of Biostatistics University of Michigan zhenkewu@umich.edu http://zhenkewu.com/teaching/graphical_model 15 September, 2016 Zhenke Wu BIOSTAT830 Graphical
More information9.5 Equivalence Relations
9.5 Equivalence Relations You know from your early study of fractions that each fraction has many equivalent forms. For example, 2, 2 4, 3 6, 2, 3 6, 5 30,... are all different ways to represent the same
More informationMini-Buckets: A General Scheme for Generating Approximations in Automated Reasoning
Mini-Buckets: A General Scheme for Generating Approximations in Automated Reasoning Rina Dechter* Department of Information and Computer Science University of California, Irvine dechter@ics. uci. edu Abstract
More informationThe Basics of Graphical Models
The Basics of Graphical Models David M. Blei Columbia University September 30, 2016 1 Introduction (These notes follow Chapter 2 of An Introduction to Probabilistic Graphical Models by Michael Jordan.
More informationPATTERN RECOGNITION AND MACHINE LEARNING CHAPTER 8: GRAPHICAL MODELS
PATTERN RECOGNITION AND MACHINE LEARNING CHAPTER 8: GRAPHICAL MODELS Bayesian Networks Directed Acyclic Graph (DAG) Bayesian Networks General Factorization Bayesian Curve Fitting (1) Polynomial Bayesian
More informationSimpler, Linear-time Transitive Orientation via Lexicographic Breadth-First Search
Simpler, Linear-time Transitive Orientation via Lexicographic Breadth-First Search Marc Tedder University of Toronto arxiv:1503.02773v1 [cs.ds] 10 Mar 2015 Abstract Comparability graphs are the undirected
More informationAbstract. 2 Background 2.1 Belief Networks. 1 Introduction
Probabilistic Partial Evaluation: Exploiting rule structure in probabilistic inference* David Poole Department of Computer Science University of British Columbia 2366 Main Mall, Vancouver, B.C., Canada
More informationFoundations of Computer Science Spring Mathematical Preliminaries
Foundations of Computer Science Spring 2017 Equivalence Relation, Recursive Definition, and Mathematical Induction Mathematical Preliminaries Mohammad Ashiqur Rahman Department of Computer Science College
More informationModule 11. Directed Graphs. Contents
Module 11 Directed Graphs Contents 11.1 Basic concepts......................... 256 Underlying graph of a digraph................ 257 Out-degrees and in-degrees.................. 258 Isomorphism..........................
More informationRepresentation of Finite Games as Network Congestion Games
Representation of Finite Games as Network Congestion Games Igal Milchtaich To cite this version: Igal Milchtaich. Representation of Finite Games as Network Congestion Games. Roberto Cominetti and Sylvain
More informationThis article was originally published in a journal published by Elsevier, and the attached copy is provided by Elsevier for the author s benefit and for the benefit of the author s institution, for non-commercial
More informationOn 2-Subcolourings of Chordal Graphs
On 2-Subcolourings of Chordal Graphs Juraj Stacho School of Computing Science, Simon Fraser University 8888 University Drive, Burnaby, B.C., Canada V5A 1S6 jstacho@cs.sfu.ca Abstract. A 2-subcolouring
More informationInformation Granulation and Approximation in a Decision-theoretic Model of Rough Sets
Information Granulation and Approximation in a Decision-theoretic Model of Rough Sets Y.Y. Yao Department of Computer Science University of Regina Regina, Saskatchewan Canada S4S 0A2 E-mail: yyao@cs.uregina.ca
More informationBayesian Network & Anomaly Detection
Study Unit 6 Bayesian Network & Anomaly Detection ANL 309 Business Analytics Applications Introduction Supervised and unsupervised methods in fraud detection Methods of Bayesian Network and Anomaly Detection
More informationLearning Equivalence Classes of Bayesian-Network Structures
Journal of Machine Learning Research 2 (2002) 445-498 Submitted 7/01; Published 2/02 Learning Equivalence Classes of Bayesian-Network Structures David Maxwell Chickering Microsoft Research One Microsoft
More informationMonotone Paths in Geometric Triangulations
Monotone Paths in Geometric Triangulations Adrian Dumitrescu Ritankar Mandal Csaba D. Tóth November 19, 2017 Abstract (I) We prove that the (maximum) number of monotone paths in a geometric triangulation
More informationTwo Comments on the Principle of Revealed Preference
Two Comments on the Principle of Revealed Preference Ariel Rubinstein School of Economics, Tel Aviv University and Department of Economics, New York University and Yuval Salant Graduate School of Business,
More informationDistributed minimum spanning tree problem
Distributed minimum spanning tree problem Juho-Kustaa Kangas 24th November 2012 Abstract Given a connected weighted undirected graph, the minimum spanning tree problem asks for a spanning subtree with
More informationSpanning tree congestion critical graphs
Spanning tree congestion critical graphs Daniel Tanner August 24, 2007 Abstract The linear or cyclic cutwidth of a graph G is the minimum congestion when G is embedded into either a path or a cycle respectively.
More informationMassachusetts Institute of Technology Department of Electrical Engineering and Computer Science Algorithms For Inference Fall 2014
Suggested Reading: Massachusetts Institute of Technology Department of Electrical Engineering and Computer Science 6.438 Algorithms For Inference Fall 2014 Probabilistic Modelling and Reasoning: The Junction
More information1 Matchings with Tutte s Theorem
1 Matchings with Tutte s Theorem Last week we saw a fairly strong necessary criterion for a graph to have a perfect matching. Today we see that this condition is in fact sufficient. Theorem 1 (Tutte, 47).
More informationLearning Bayesian Networks via Edge Walks on DAG Associahedra
Learning Bayesian Networks via Edge Walks on DAG Associahedra Liam Solus Based on work with Lenka Matejovicova, Adityanarayanan Radhakrishnan, Caroline Uhler, and Yuhao Wang KTH Royal Institute of Technology
More informationSeparators and Adjustment Sets in Markov Equivalent DAGs
Proceedings of the Thirtieth AAAI Conference on Artificial Intelligence (AAAI-16) Separators and Adjustment Sets in Markov Equivalent DAGs Benito van der Zander and Maciej Liśkiewicz Institute of Theoretical
More informationSchema Design for Uncertain Databases
Schema Design for Uncertain Databases Anish Das Sarma, Jeffrey Ullman, Jennifer Widom {anish,ullman,widom}@cs.stanford.edu Stanford University Abstract. We address schema design in uncertain databases.
More informationJoint Entity Resolution
Joint Entity Resolution Steven Euijong Whang, Hector Garcia-Molina Computer Science Department, Stanford University 353 Serra Mall, Stanford, CA 94305, USA {swhang, hector}@cs.stanford.edu No Institute
More information2386 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 52, NO. 6, JUNE 2006
2386 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 52, NO. 6, JUNE 2006 The Encoding Complexity of Network Coding Michael Langberg, Member, IEEE, Alexander Sprintson, Member, IEEE, and Jehoshua Bruck,
More informationThroughout the chapter, we will assume that the reader is familiar with the basics of phylogenetic trees.
Chapter 7 SUPERTREE ALGORITHMS FOR NESTED TAXA Philip Daniel and Charles Semple Abstract: Keywords: Most supertree algorithms combine collections of rooted phylogenetic trees with overlapping leaf sets
More informationCausal Modeling of Observational Cost Data: A Ground-Breaking use of Directed Acyclic Graphs
use Causal Modeling of Observational Cost Data: A Ground-Breaking use of Directed Acyclic Graphs Bob Stoddard Mike Konrad SEMA SEMA November 17, 2015 Public Release; Distribution is Copyright 2015 Carnegie
More informationTHE TRANSITIVE REDUCTION OF A DIRECTED GRAPH*
SIAM J. COMPUT. Vol. 1, No. 2, June 1972 THE TRANSITIVE REDUCTION OF A DIRECTED GRAPH* A. V. AHO, M. R. GAREY" AND J. D. ULLMAN Abstract. We consider economical representations for the path information
More informationFOUR EDGE-INDEPENDENT SPANNING TREES 1
FOUR EDGE-INDEPENDENT SPANNING TREES 1 Alexander Hoyer and Robin Thomas School of Mathematics Georgia Institute of Technology Atlanta, Georgia 30332-0160, USA ABSTRACT We prove an ear-decomposition theorem
More informationThe Encoding Complexity of Network Coding
The Encoding Complexity of Network Coding Michael Langberg Alexander Sprintson Jehoshua Bruck California Institute of Technology Email: mikel,spalex,bruck @caltech.edu Abstract In the multicast network
More informationLoopy Belief Propagation
Loopy Belief Propagation Research Exam Kristin Branson September 29, 2003 Loopy Belief Propagation p.1/73 Problem Formalization Reasoning about any real-world problem requires assumptions about the structure
More informationCounting and Exploring Sizes of Markov Equivalence Classes of Directed Acyclic Graphs
Journal of Machine Learning Research 16 (2015) 2589-2609 Submitted 9/14; Revised 3/15; Published 12/15 Counting and Exploring Sizes of Markov Equivalence Classes of Directed Acyclic Graphs Yangbo He heyb@pku.edu.cn
More informationMultiple Constraint Satisfaction by Belief Propagation: An Example Using Sudoku
Multiple Constraint Satisfaction by Belief Propagation: An Example Using Sudoku Todd K. Moon and Jacob H. Gunther Utah State University Abstract The popular Sudoku puzzle bears structural resemblance to
More informationAn efficient approach for finding the MPE in belief networks
342 Li and D'Ambrosio An efficient approach for finding the MPE in belief networks Zhaoyu Li Department of Computer Science Oregon State University Corvallis, OR 97331 Bruce D'Ambrosio Department of Computer
More informationChecks and Balances - Constraint Solving without Surprises in Object-Constraint Programming Languages: Full Formal Development
Checks and Balances - Constraint Solving without Surprises in Object-Constraint Programming Languages: Full Formal Development Tim Felgentreff, Todd Millstein, Alan Borning and Robert Hirschfeld Viewpoints
More informationA Model of Machine Learning Based on User Preference of Attributes
1 A Model of Machine Learning Based on User Preference of Attributes Yiyu Yao 1, Yan Zhao 1, Jue Wang 2 and Suqing Han 2 1 Department of Computer Science, University of Regina, Regina, Saskatchewan, Canada
More informationImproved upper and lower bounds on the feedback vertex numbers of grids and butterflies
Improved upper and lower bounds on the feedback vertex numbers of grids and butterflies Florent R. Madelaine and Iain A. Stewart Department of Computer Science, Durham University, Science Labs, South Road,
More informationA CSP Search Algorithm with Reduced Branching Factor
A CSP Search Algorithm with Reduced Branching Factor Igor Razgon and Amnon Meisels Department of Computer Science, Ben-Gurion University of the Negev, Beer-Sheva, 84-105, Israel {irazgon,am}@cs.bgu.ac.il
More informationChapter 21: Metatheory of Recursive Types. Induction and Coinduction Finite and Infinite Types/Subtyping Membership Checking
Chapter 21: Metatheory of Recursive Types Induction and Coinduction Finite and Infinite Types/Subtyping Membership Checking Review of Chapter 20 Recursive Types Lists NatList =
More informationMachine Learning. Sourangshu Bhattacharya
Machine Learning Sourangshu Bhattacharya Bayesian Networks Directed Acyclic Graph (DAG) Bayesian Networks General Factorization Curve Fitting Re-visited Maximum Likelihood Determine by minimizing sum-of-squares
More informationProbabilistic Partial Evaluation: Exploiting rule structure in probabilistic inference
Probabilistic Partial Evaluation: Exploiting rule structure in probabilistic inference David Poole Department of Computer Science University of British Columbia 2366 Main Mall, Vancouver, B.C., Canada
More informationRecognizing Interval Bigraphs by Forbidden Patterns
Recognizing Interval Bigraphs by Forbidden Patterns Arash Rafiey Simon Fraser University, Vancouver, Canada, and Indiana State University, IN, USA arashr@sfu.ca, arash.rafiey@indstate.edu Abstract Let
More informationCS2 Algorithms and Data Structures Note 10. Depth-First Search and Topological Sorting
CS2 Algorithms and Data Structures Note 10 Depth-First Search and Topological Sorting In this lecture, we will analyse the running time of DFS and discuss a few applications. 10.1 A recursive implementation
More informationCausal Explanation with Background Knowledge. Bhaskara Reddy Moole
Causal Explanation with Background Knowledge Bhaskara Reddy Moole bhaskarareddy@wondertechnology.com School of Management Walden University, 155 Fifth Ave South Minneapolis, MN 55401. Marco Valtorta mgv@cse.sc.edu
More informationA Decision-Theoretic Rough Set Model
A Decision-Theoretic Rough Set Model Yiyu Yao and Jingtao Yao Department of Computer Science University of Regina Regina, Saskatchewan, Canada S4S 0A2 {yyao,jtyao}@cs.uregina.ca Special Thanks to Professor
More informationA GRAPH FROM THE VIEWPOINT OF ALGEBRAIC TOPOLOGY
A GRAPH FROM THE VIEWPOINT OF ALGEBRAIC TOPOLOGY KARL L. STRATOS Abstract. The conventional method of describing a graph as a pair (V, E), where V and E repectively denote the sets of vertices and edges,
More informationK s,t -saturated bipartite graphs
K s,t -saturated bipartite graphs Wenying Gan Dániel Korándi Benny Sudakov Abstract An n-by-n bipartite graph is H-saturated if the addition of any missing edge between its two parts creates a new copy
More informationCS 343: Artificial Intelligence
CS 343: Artificial Intelligence Bayes Nets: Independence Prof. Scott Niekum The University of Texas at Austin [These slides based on those of Dan Klein and Pieter Abbeel for CS188 Intro to AI at UC Berkeley.
More informationEXTREME POINTS AND AFFINE EQUIVALENCE
EXTREME POINTS AND AFFINE EQUIVALENCE The purpose of this note is to use the notions of extreme points and affine transformations which are studied in the file affine-convex.pdf to prove that certain standard
More informationBayesian Networks. 2. Separation in Graphs
ayesian Netwks ayesian Netwks 2. Separation in raphs Lars Schmidt-Thieme Infmation Systems and Machine Learning Lab (ISMLL) Institute f usiness conomics and Infmation Systems & Institute f omputer Science
More informationBasics of Graph Theory
Basics of Graph Theory 1 Basic notions A simple graph G = (V, E) consists of V, a nonempty set of vertices, and E, a set of unordered pairs of distinct elements of V called edges. Simple graphs have their
More informationStar coloring planar graphs from small lists
Star coloring planar graphs from small lists André Kündgen Craig Timmons June 4, 2008 Abstract A star coloring of a graph is a proper vertex-coloring such that no path on four vertices is 2-colored. We
More informationByzantine Consensus in Directed Graphs
Byzantine Consensus in Directed Graphs Lewis Tseng 1,3, and Nitin Vaidya 2,3 1 Department of Computer Science, 2 Department of Electrical and Computer Engineering, and 3 Coordinated Science Laboratory
More information3 Fractional Ramsey Numbers
27 3 Fractional Ramsey Numbers Since the definition of Ramsey numbers makes use of the clique number of graphs, we may define fractional Ramsey numbers simply by substituting fractional clique number into
More informationPartitioning Complete Multipartite Graphs by Monochromatic Trees
Partitioning Complete Multipartite Graphs by Monochromatic Trees Atsushi Kaneko, M.Kano 1 and Kazuhiro Suzuki 1 1 Department of Computer and Information Sciences Ibaraki University, Hitachi 316-8511 Japan
More informationDischarging and reducible configurations
Discharging and reducible configurations Zdeněk Dvořák March 24, 2018 Suppose we want to show that graphs from some hereditary class G are k- colorable. Clearly, we can restrict our attention to graphs
More information